Weakly-Supervised Shape Multi-Completion of Point Clouds by Structural Decomposition

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geometric Deep Learning

Background:

• Partial point clouds present significant challenges for complete 3D mesh generation.
• Existing methods struggle with data accessibility, shape preservation, and robustness on real-scan data.
• Signed Distance Functions (SDFs) are often required during training, limiting flexibility.

Purpose of the Study:

• To develop an innovative weakly-supervised shape completion method for 3D meshes.
• To overcome limitations of current methods by leveraging structural information.
• To eliminate the need for SDFs during training.

Main Methods:

• A weakly-supervised shape completion method using structural decomposition.
• Representing objects as abstract structural frameworks and part details.
• Utilizing a completion network for image-based part details and a diffusion-based network for multi-result generation.

Main Results:

• Achieved state-of-the-art (SOTA) performance in 3D shape completion.
• Demonstrated an average improvement of over 38.1% compared to prior methods.
• Showcased superior robustness and accuracy on both artificial and real-scan datasets.

Conclusions:

• The proposed method effectively generates complete 3D meshes from partial point clouds.
• Weakly-supervised learning via structural decomposition offers a robust alternative to SDF-based training.
• The approach significantly advances the field of 3D shape completion.